lixiaoxiao9888-create/manju-laoli-skill

---

🏆 赞助商

感谢以下赞助商对本项目的支持：

💡 招商说明：赞助位对模型渠道、算力平台、API 服务商及创作团队开放。 合作请联系：邮箱 [email protected] · 微信 lisenlin998

🎬 分镜引擎专项赞助用户

感谢以下用户赞助分镜引擎的开发（开放中）：

你的头像将出现在这里。

---

🧭 这是什么

一套 给 AI 助手执行的 AIGC 全流程创作规则库（OpenClaw Skill / Agent 技能包）。V7.4.20 起升级为「多技能套装」：1 个总控（short-drama-director）+ 12 个子技能（sd-*-engine），四层纵深架构——不是单一 SKILL.md，而是一套可路由、可挂载、可机检的导演系统。

核心理念：「用户看到的是 5 个步骤，AI 内部跑的是 100 条规则」——用户层只做「① 做什么 ② 从哪来 ③ 模型 ④ 画幅 ⑤ 要什么结果」，锁 / 路由 / 子技能挂载 / 机检全部由总控内部完成，用户不需要手动调用任何子技能（直接说需求，总控自动路由；但六道锁必须弹出让用户勾选）。

它把「AIGC 视听创作」拆成一条可复用的工业化管线，核心解决三件事：

1. 不漂移 —— 角色 / 场景 / 道具先锁定再进分镜（Asset-First），跨镜跨集一致性有据可依；
2. 不返工 —— 六道前置锁 + 投喂必查清单 + P0~P2 独立门禁 + 稿件级机检，把错误拦在生成之前；
3. 能直接投喂 —— 输出 七段式结构化提示词，整组复制即可喂给视频模型；走 MiniMax H3 时自动转 H3 原生六段式。

🔑 本包不持有图：交付的是「提示词 + 接线说明书」——图由用户在外部平台（即梦 / Seedance / 豆包等）自行生成并持有。素材按存在形态分三档：M0 纯提示词 / M1 编号占位（默认）/ M2 已入库，每次交付首行声明形态。

🎯 适用场景

管线与题材解耦，凡「剧本 → 分镜 → 视频模型投喂」的 AIGC 创作都可套用：

• AI 短剧 / 漫剧 / 短视频剧集（本包最典型的高密度场景）
• AI 广告片 / 品牌宣传片 / 概念片
• 分镜预演（PREVIS）与动态分镜 · 武打 / 动作场面设计（现代格斗、武侠、仙侠、二次元、科幻）
• 历史 / 科普 / 盘点解说叙事（赛道③，挂载 sd-jieshuo-engine）
• 海外剧 / 多语 / 方言（赛道②，挂载 sd-overseas-engine）
• 情绪戏 / 微表情专项（FACS 表情工程）
• 任何需要跨镜头角色一致性的 AI 视频生产

🧩 四层纵深架构（1 总控 + 12 子技能）

🔒 六道前置锁 + I 画风锁（动手前全部弹出让用户勾选）

⚠️ 任何一道未指定，必须停下询问并把选项摆出来让用户勾——禁止默认跳过、禁止替用户推断（硬门，不是建议）。另含动作强度三档定档：R1 写实克制 / R2 商业高燃（默认） / R3 玄幻大招。

🔥 核心能力

🖥️ 适用平台与模型

---

⚙️ 快速开始

📦 部署（Multi-Agent 套装）：本仓根目录并列 1 个总控（short-drama-director/）+ 12 个子技能（sd-*-engine/）。支持多技能目录的平台（如 OpenClaw）把 13 个技能目录整体同级部署即可，主入口 = short-drama-director；要求「zip 根即 SKILL.md」的单技能平台，请以 short-drama-director/SKILL.md 为主入口并把子技能同级补齐。

# 安装到 OpenClaw（总控 + 12 子技能整体同级部署）
openclaw skills install ./short-drama-director --as short-drama-director

# 包级自检（50 份规则库静态 + 语义回归）
python3 short-drama-director/scripts/check_package.py

# 出稿机检（FAIL 清零才交付；--model 2.5|2.0|h3，--track 1~4 按赛道分流）
python3 short-drama-director/scripts/validate_prompt.py .md --model 2.5

# 投喂层六道闸门（含素材形态一致性）
python3 sd-feed-engine/scripts/gate_check.py .md

# 项目 md → 单文件离线看板
python3 short-drama-director/scripts/build_board_lite.py 

给 AI 助手的最小执行序列：读 SKILL.md → 完成六道锁（〇·0 赛道 / 〇·1 素材来源 / A 模型 / B 画幅 / C 交付形态 / D 段缝）→ 定动作强度档（R1/R2/R3）→ 有对白先过语速自检 + 语气位 → 资产锁定 → 分镜 → 按 ★ prompt-feeding-checklist.md 生成投喂词 → 跑 validate_prompt.py 机检 → 过 quality-gate-review。

🧮 常用指令

/写剧本 · /台词诊断 · /数字资产包 · /素材形态 · /接线单 · /做分镜 · /语速自检 · /生成视频提示词 · /H3提示词 · /导出工作流参数 · /情绪曲线 · /看板 · /微表情 · /审查 · 跳过确认，直接出整集

完整指令路由见 short-drama-director/SKILL.md

---

📦 项目结构

lixiaoxiao9888-create/manju-laoli-skill/        # V7.4.26 多技能套装（164 文件）
├── README.md                                   # 本落地页（中文 + English）
├── LICENSE / .gitignore
├── short-drama-director/                       # L0 总控（62 文件 · 50 份规则库 · 机检脚本）
│   ├── SKILL.md                                #   总控路由 + 工作法 + 模块索引
│   ├── references/                             #   48 份专业规则库（含 style-bible 画风锁知识库 10 份）
│   ├── scripts/                                #   check_package / validate_prompt / gate 等机检全家桶
│   └── README.md / LICENSE / CHANGELOG.md
├── sd-decision-engine/                         # L1 决策层（情境映射 E01~E15）
├── sd-recon-engine/                            # L2 生产层 · 素材采风
├── sd-script-engine/                           # L2 · 剧本（国内剧 + 外来剧本润色）
├── sd-overseas-engine/                         # L2 · 海外剧剧本（赛道②）
├── sd-jieshuo-engine/                          # L2 · 解说叙事剧本（赛道③）
├── sd-asset-engine/                            # L2 · 资产锁定
├── sd-storyboard-engine/                       # L2 · 文戏分镜 + 故事板图
├── sd-combat-engine/                           # L2 · 通用打戏
├── sd-xuanhuan-engine/                         # L2 · 玄幻仙侠打戏
├── sd-real-engine/                             # L2 · 真人写实打戏
├── sd-feed-engine/                             # L3 交付层 · 投喂（六道闸门）
└── sd-qc-engine/                               # L3 · 质检 / 连续性 / 声音设计

---

📜 更新历史

• V7.4.26（2026-10-21）— 审计补丁 P1 + P4.5 交付标配/机检：修 9 处执行链缺陷——① P4.5 触发条件机检化（契约 CT5-P45：≥3 人同框 / ≥2 个 @//scn/ 场景 / 单段跨 2 场，命中不附即 FAIL），本仓再升为默认必出；② 新增 P3/P4 最小必读集（治 12.1 万 token vs 8k 预算）；③ 七段式镜行升级五要素（景别・焦段・机位角度・运镜（用途标签）），资产引用升级 @//scn/ @//chr/ @//prp/（旧写法兼容）；④ C3 升 FAIL（段时码必须从 00:00 起）、新增 C18g2（缺焦段）、--by-segment 改默认、C11b/C18m 加 D 锁②硬切豁免、C12 版式豁免、补 9 个运镜词、C17c/C26b/C5/C29/C19c 误报修复、C31 区分项目包/投喂单元；⑤ 新增 check_p45.py（P450~P458）结构机检。创作口径（锁 / 七段式 / 画风 / 打戏）未动。
• V7.4.25（2026-10-19）— 交付层与门禁实战回写：以《兽瞳》第01集全流程实跑为样本，修四类反复踩到的坑——① 交付件数失控（单集曾产出 15 文件/10 目录）：定死「对外最多 3 件」（主文档 / 看板 / P4.5），md 正文只放资产提示词 + 分镜提示词，其余归 _内部_ 前缀的内部档案，命名全中文；② 资产提示词结构性缺陷：新增 scripts/check_asset_blocks.py 门禁（A1 跨条引用「同上/同规格/见母版/同图N」、A2 16:9、A3 九宫格八字段、A4·A5 角色四格与「主图」旧称致 5 格、A6 道具四格字段名），P2 出稿 FAIL 清零；③ 机检口径失真：validate_prompt.py 新增 --by-segment，多段稿逐段独立判据——整份送检会把 3×30s 累计成 90s 一段，误判打戏档触发 C21/C22 并让 C10/C15 失真；④ 新增 references/delivery-pitfalls.md（交付翻车清单）与 P4「段级红线八条」（特写≥25% / 单镜≤6s / 机位不重复 / C32 背景光线写在台词前 / 运镜词表 / 人数肯定句 / 段内镜号从镜1 起 / 段缝换景别）。创作知识层（剧本、表演、打戏、画风）未改动。
• V7.4.23（2026-10-16）— 粉丝排查报告核验 + 全包「文档法定词 × 机检词表」审计全修：粉丝（Claude / Codex）报告指 camera-specs-15rules.md §二 法定运镜「命中瞬间定格（Hit-Stop）」未入机检词表 CAM_FAMILIES → 照文档写必报 C18g 假错，且为规避该误报改写成「固定」的执行者又被 C22c 报"缺顿"（两门夹杀、无合规写法）。逐条端到端复现，属实。核验中把该根因推广到全包反向核对，另查出 25 个词解析为空集（定格族 11 词 + 环移 + 模式F 运镜库 7 词 + 焦点纵深拉移…），并修 2 处 V7.4.21 本轮引入的缺陷：① C32 与「光影单一基准」教义互锁夹杀（教义要求「时段光影立【场景资产】为唯一基准、镜行只写增量」，C32 却逐镜索要光线方向）→ 改段级继承口径（段内有基准则逐镜只判背景）；② BG_ELEM_RE 过窄（「脚下泥土犁出两道浅沟，竹叶被气浪掀飞」被漏判成"无背景"）→ 补自然环境元素 + 环境行为词。另修 model-adapters.md 两节重复编号 5.0.5（光影单一基准改 §5.0.5b）、新增 ◑17 全包零语法告警门。⚠️ V7.4.21 的交付 zip 请勿推送（含上述 C32 互锁），以 V7.4.25 为准
• V7.4.21（2026-10-16）— 粉丝「森林」视感反馈全修（4 条 + 追加 3 条，逐条核验真伪后才动手）：新增 §七 四轴联动矩阵（景别×焦段×机位×运镜 8 档一次性写齐 + 3 条组合级禁令 + 自检四问）· §一 景别占比预算表（六组 × 文戏/打戏分列，补「景别维度只有特写一个区间」缺口）· §二 景别→焦段逐级对照表（补齐中近景 / 近景空档）· §四 变焦族四件套（推拉镜 / 光学变焦 Zoom / 急速变焦 Snap Zoom / 滑动变焦 Dolly Zoom 四条独立成档，并修机检别名误映射——Zoom 曾被错归滑动变焦且被吞进「全片 ≤1」名额）；资产体系新增第四族「群像资产卡 GRP-N」（7 字段 + 3 格出图 + @群体名 锚点，与 @人名 同栏同待遇）；画面段补最低三要素（人物位置/姿态 + 背景元素 + 光线方向）。机检：子检查码 98 → 99（新增 C18p 景别占比越界），增量主码 C27~C32（新增 C32 画面段最低三要素；编号避让——C18o / C30 已被既有占用，故取 C18p / C32）；并根治 check_package.py 扫描域（收敛为 suite_roots() 恰 13 目录 + 新增 ◑16 扫描域锁定断言，不再被仓库根落地页或同级其它技能误伤假 FAIL）
• V7.4.20 多技能套装版（2026-10-15）— 架构升级为「多技能套装」：1 总控 + 12 子技能 · 四层纵深（L0 总控 / L1 决策 / L2 生产 / L3 交付）；规则库扩至 48 份（含 style-bible/ 画风锁知识库 10 份），全包 154 文件；〇·0 赛道分流（含赛道② 海外剧 · 赛道③ 解说叙事）与 〇·1 素材来源（对标采风）正交可组合；素材形态 M0/M1/M2 +《素材接线单》（本包不持有图）；新增 ★场面调度（blocking-direction）· ★故事板图·面板链 · 剧情分镜格 · 场景空间事实五件事；机检扩至 C1~C26（子检查码 98 条，另设增量主码 C27 对话戏节奏 / C28 台词机械化 / C29 资产状态偏离 / C30 外挂指纹与交接探针）、H3 支线 H1~H12；投喂层新增六道闸门（含素材形态一致性）
• V6.9.8 轻量版（2026-09-20）— 新增 MiniMax H3 格式支线：A 锁并列第三选项，只做「七段式 → H3 原生格式」转换（h3-adapter.md：两 checkpoint 路由 / 六段式映射 / 素材上限 9 图·3 视频·3 音频·总 12 / 9 张取材策略 / 两套投喂载荷）；出稿机检新增 H1~H12（--model h3）
• V6.9.3 ~ V6.9.7 轻量版（2026-09-19）— 五轮实拍复盘驱动的定向修订：对白语气位（C16）· 人物账连续·摆位必入画（C17）· 首镜锚定·景别一致·姿态延续·视线指派（C18）· 禁令失效三改法 + 栏目纪律（C19 / C19e）· 首镜单主体锚定与具象化（C20）· 示范句验证纪律；机检升至 C1~C20
• V6.9 轻量版（2026-09-18）— 新增 〇节 D·段缝衔接方式三选一锁 / 旁白·独白视听编织规程（OS 声画分离红线 · VO 冷漠审视）/ 时长守恒·禁拉伸凑满
• V6.8 轻量版（2026-09-15）— 可灵移出正式支持（降级 LEGACY）/ 角色资产统一 4 View / 空间 S0~S4 · 资产 A0~A3 · FACS 三级按需触发 / 2.5 新增任务模式二选一
• V6.7 轻量版（2026-09-14）— 双锁同规格（模型版本 + 画幅三选一强制锁）/ 按官方文档校准 Seedance 2.5 / 吸收 FACS 微表情引擎三件套 / 吸收轻量离线分镜看板
• V6.5 Multi-Agent（2026-09-01）— Asset-First 六阶段 + 多 Agent 平台适配 + Seedance 2.5 深度适配 + 静止段禁用铁律
• 旧版归档：manju-laoli-skill-legacy

---

---

🧭 What This Is

An AIGC creation rulebase for AI agents (OpenClaw Skill / Agent skill package). As of V7.4.20 it is a multi-skill suite — 1 router (short-drama-director) + 12 sub-skills (sd-*-engine) across four layers.

Core idea: “The user sees 5 steps; the AI runs 100 rules.” The user only decides ① what to make ② where it comes from ③ model ④ aspect ratio ⑤ desired output — locks, routing, sub-skill mounting, and linting are all handled internally. Sub-skills are never called by hand.

It turns AIGC audiovisual creation into a reusable industrial pipeline, solving three things:

1. No drift — characters / scenes / props are locked first, then storyboarded (Asset-First);
2. No rework — six front-end locks + a pre-feed checklist + independent P0–P2 gates + shot-level lint stop errors before generation;
3. Feed-ready — outputs a structured 7-section prompt you can copy straight into the video model; MiniMax H3 selection auto-converts it into H3’s native 6-section format.

🔑 The package holds no images: it delivers “prompts + a wiring sheet”. Images are generated and held by the user on external platforms. Assets are graded by existence form — M0 prompt-only / M1 numbered slots (default) / M2 registered — declared on the first line of every delivery.

🎯 Use Cases

Genre-agnostic: anything that goes script → storyboard → video-model feed fits.

• AI short dramas / animated shorts / short-form series (the densest use case)
• AI commercials / brand films / concept films
• Storyboard previsualization (PREVIS) · action & fight choreography design
• History / science explainer narration (track ③, sd-jieshuo-engine)
• Overseas drama / multi-language / dialect (track ②, sd-overseas-engine)
• Emotional & micro-expression work (FACS expression engineering)
• Any AI video production needing cross-shot character consistency

🧩 Four-Layer Architecture (1 router + 12 sub-skills)

🔒 Six Front-End Locks + Style Lock (all must be offered for the user to choose)

⚠️ If any lock is unspecified, the system must stop and offer the options — no skipping, no inferring. Action-intensity tiers: R1 restrained / R2 commercial (default) / R3 fantasy.

🔥 Capabilities

🖥️ Platforms & Models

---

⚙️ Quick Start

📦 Deploy (multi-agent suite): the repo root holds 1 router (short-drama-director/) + 12 sub-skills (sd-*-engine/) side by side. On multi-skill platforms (e.g. OpenClaw) deploy all 13 directories at the same level; the main entry is short-drama-director.

# Install into OpenClaw
openclaw skills install ./short-drama-director --as short-drama-director

# Package lint (48 rulebooks, static + semantic regression)
python3 short-drama-director/scripts/check_package.py

# Shot-level feed lint (zero FAIL to ship; --model 2.5|2.0|h3, --track 1~4)
python3 short-drama-director/scripts/validate_prompt.py .md --model 2.5

# Feed-layer six gates (incl. asset-form consistency)
python3 sd-feed-engine/scripts/gate_check.py .md

# Project md → single-file offline board
python3 short-drama-director/scripts/build_board_lite.py 

Minimal agent sequence: read SKILL.md → complete the six locks → pick action tier (R1/R2/R3) → dialogue speed check + tone slot → lock assets → storyboard → generate feed prompts → pass validate_prompt.py → pass quality-gate-review.

📦 Structure

lixiaoxiao9888-create/manju-laoli-skill/        # V7.4.26 multi-skill suite (164 files)
├── README.md                                   # landing (Chinese + English)
├── LICENSE / .gitignore
├── short-drama-director/                       # L0 router (62 files · 48 rulebooks · lint scripts)
├── sd-decision-engine/                         # L1 decision (E01–E15)
├── sd-recon-engine/                            # L2 · sourcing
├── sd-script-engine/                           # L2 · scriptwriting
├── sd-overseas-engine/                         # L2 · overseas script (track ②)
├── sd-jieshuo-engine/                          # L2 · explainer script (track ③)
├── sd-asset-engine/                            # L2 · asset locking
├── sd-storyboard-engine/                       # L2 · storyboard + storyboard sheets
├── sd-combat-engine/                           # L2 · general fight
├── sd-xuanhuan-engine/                         # L2 · xianxia fight
├── sd-real-engine/                             # L2 · live-action fight
├── sd-feed-engine/                             # L3 delivery · feed (six gates)
└── sd-qc-engine/                               # L3 · QC / continuity / sound

---

📜 Changelog

• V7.4.26 (2026-10-21) — audit patch P1 + P4.5 delivery standard/check: fixes nine execution-chain defects — ① P4.5 triggers are now machine-checked (contract CT5-P45: ≥3 named subjects in one stance block / ≥2 distinct @//scn/ scenes / a segment spanning 2 scenes; hit without attachment = FAIL), and this repo further makes it always emitted; ② new P3/P4 minimum reading sets (against the 121k-token vs 8k budget gap); ③ the 7-section shot line gains five elements (shot size · focal length · camera angle · movement with usage tag) and asset refs upgrade to @//scn/ @//chr/ @//prp/; ④ C3 becomes FAIL (segment timecodes must start at 00:00), new C18g2 (missing focal length), --by-segment becomes the default, C11b/C18m gain D-lock-② hard-cut exemptions, C12 layout exemptions, nine movement words added, false-positive fixes for C17c/C26b/C5/C29/C19c, and C31 now distinguishes project packages from feed units; ⑤ new check_p45.py (P450–P458) structural gate. Creative doctrine unchanged.
• V7.4.25 (2026-10-19) — delivery-layer and gate hardening, written back from a real run: based on a full pipeline run of episode 1, four recurring failure classes are fixed — ① deliverable explosion (one episode once produced 15 files / 10 folders): hard cap of three external artifacts (main doc / board / P4.5), the md body keeps only asset prompts and shot prompts, everything else moves to _内部_ internal archives, and all names are localized; ② structural asset-prompt defects: new scripts/check_asset_blocks.py gate (A1 cross-references such as “same as above / see master / same as image N”, A2 16:9, A3 the nine-grid eight fields, A4/A5 four-panel character boards and the retired “main image” wording that yields five panels, A6 prop four-panel field names), FAIL-free before P3; ③ validation granularity: validate_prompt.py --by-segment validates each 30s segment independently — whole-file parsing stacks 3x30s into one 90s block, mis-classifies it as a combat script (C21/C22) and skews C10/C15; ④ new references/delivery-pitfalls.md plus eight per-segment red lines in P4. The creative knowledge layer is unchanged.
• V7.4.23 (2026-10-16) — fan-report verification + suite-wide “doc-legal-word × lint-table” audit pass: a fan report (Claude / Codex) showed that the camera-movement word 「命中瞬间定格」/ Hit-Stop — legalized by camera-specs-15rules.md §2 — was absent from the lint table CAM_FAMILIES, so writing it exactly as the docs prescribe raised a false C18g, while the workaround (rewriting it as 「固定」/ static) got flagged by C22c for “missing beat” — a two-gate squeeze with no compliant form. Reproduced end-to-end; confirmed real. Generalizing the root cause across the package surfaced 25 more words that resolve to an empty family (11 hit-stop words + 环移 + 7 mode-F library words + focus-pull …), and fixed 2 defects introduced by V7.4.21 itself: ① C32 vs. the single-light-baseline doctrine (the doctrine says the segment’s 时段光影 is the only baseline and shot lines carry deltas only, while C32 demanded light direction per shot) → switched to a segment-level inheritance rule; ② BG_ELEM_RE too narrow (“脚下泥土犁出两道浅沟，竹叶被气浪掀飞” judged as no background) → widened with natural elements and environment-action verbs. Also fixed the duplicate heading 5.0.5 in model-adapters.md (light baseline → §5.0.5b) and added a ◑17 zero-syntax-warning gate. ⚠️ Do not push the V7.4.21 zip (it carries the C32 mutual lock); ship V7.4.23 instead
• V7.4.21 (2026-10-16) — full fix pass on fan 「森林」's visual feedback (4 items + 3 follow-ups, each verified before acting): new §7 four-axis linkage matrix (shot size × focal length × camera position × movement written out in 8 tiers + 3 combination-level bans + a 4-question self-check) · §1 shot-size budget table (6 groups × drama/fight, filling the “shot-size dimension had only the close-up band” gap) · §2 shot-size → focal-length ladder (fills the medium-close / close gaps) · §4 zoom-family quartet (push/pull · optical Zoom · snap zoom · dolly zoom as four separate entries, fixing the lint alias mis-map where Zoom was misfiled as dolly zoom and swallowed by its “≤1 per film” quota); asset system gains a fourth family — crowd asset card GRP-N (7 fields + 3-grid sheet + @crowd-name anchor, same slot as @name); the picture segment now requires three minimum elements (subject position/pose + background element + light direction). Lint: sub-codes 98 → 99 (new C18p shot-size ratio out-of-range), incremental codes C27–C32 (new C32 picture-segment minimums; numbering avoids taken slots — C18o / C30 were occupied, hence C18p / C32); check_package.py scan scope fixed (narrowed to suite_roots() = exactly 13 dirs + new ◑16 scope-lock assertion, so the repo-root landing page or sibling skills no longer raise false FAILs)
• V7.4.20 Multi-Skill Suite (2026-10-15) — architecture upgraded to a multi-skill suite: 1 router + 12 sub-skills across four layers; rulebooks grow to 48 (incl. 10 in style-bible/), 154 files total; 〇·0 track routing (overseas ② · explainer ③) and 〇·1 sourcing (benchmark recon) orthogonal; asset existence form M0/M1/M2 + wiring sheet (no images held); new ★blocking direction · ★storyboard sheets · plot-deduction grid · scene-spatial facts; lint extended to C1~C26 (98 sub-codes; incremental C27–C30) and H3 H1–H12; feed layer gains six gates (incl. asset-form consistency)
• V6.9.8 Lite (2026-09-20) — adds the MiniMax H3 format branch; shot-level lint gains H1–H12 (--model h3)
• V6.9.3 – V6.9.7 Lite (2026-09-19) — five rounds of retro-driven revisions; lint raised to C1–C20
• V6.9 Lite (2026-09-18) — D-lock three-way seam joining · narration/inner-monologue weaving · duration conservation
• V6.8 Lite (2026-09-15) — Kling demoted to LEGACY / 4-View assets / S0–S4 · A0–A3 · FACS on-demand / 2.5 task modes
• V6.7 Lite (2026-09-14) — equal-weight dual locks / Seedance 2.5 calibrated / FACS + offline board absorbed
• V6.5 Multi-Agent (2026-09-01) — Asset-First 6 stages + multi-agent adapters
• Legacy archive: manju-laoli-skill-legacy

---

---

🙌 贡献者与团队

感谢参与产品设计、开发、测试、内容与社区建设的成员：

V7.4.25 稳定性能力与边界

新增完整章节分页、JSON/text 统一门禁、多对白累计及 JSON 单源导出。使用单源项目时必读 short-drama-director/references/project-bundle-contract.md。导出仅含资产提示词与 P3CLEAN_GRID 计划；不是已生成图片，也不是已经过 P4 模型适配的投喂稿。旧示例回归和未解决项见包外交付说明，不以机检 PASS 代替语义与实拍验收。